It seems to me that the HistogramObserver implementation has a bug. In L 1181 the forward pass is as follows:
combined_histogram = self._combine_histograms(
combined_histogram,
self.histogram,
self.upsample_rate,
downsample_rate,
start_idx,
self.bins,
)
Where _combine_histograms is defined as:
def _combine_histograms(
self,
orig_hist: torch.Tensor,
new_hist: torch.Tensor,
upsample_rate: int,
downsample_rate: int,
start_idx: int,
Nbins: int,
) -> torch.Tensor:
Therefore, combined_histogram gets mapped to orig_hist while self.histogram gets mapped to new_hist. It seems to me like it should be the other way around given the variable names.
Can someone clarify?
Thank you!!